feat: add car-mirror-symmetry example

[TMHSDigital]

Summary

Adds examples/car-mirror-symmetry/ — a generic hatchback built as one
half (a loft of 14 stations × 9-point half-rings from bottom centerline to
roof centerline) and completed by a Mirror modifier, evaluated through
the depsgraph. Wheels and lamps are separate objects mirrored the idiomatic
way: object origin parked on the symmetry plane, mesh data offset —
mirror mirrors about the object origin, so you offset the data, never the
object. The recognizable object is the contract: a car is the thing
everyone knows is bilaterally symmetric, and its failure is unmissable.

Contract witnessed

Original datablock keeps only the authored half; the depsgraph carries the
mirrored, welded whole. Closed forms, exit non-zero on failure: half
126/231/106 with 28 centerline verts · evaluated exactly 2n − c = 224
(434 edges / 212 faces), watertight, Euler 2 · every evaluated vertex
has an exact negated-X partner · bbox min.x == −max.x · wheels (96/81 →
192/162) and lamps (8/6 → 16/12) double across the plane with origins at
x == 0.

Measured check values (live runs, both binaries)

Binaries: .scratch/blender-5.1.2-windows-x64/blender.exe (reports Blender 5.1.2) and .scratch/blender-4.5.11-windows-x64/blender.exe (reports
Blender 4.5.11 LTS). Output byte-identical on both:

• body half=126/231/106, centerline=28 · eval=224/434/212, euler=2, manifold
• sym_dev 0.000e+00 (tol 2.0e-05), bbox_asym 0.000e+00
• all four mirrored parts sym_dev 0.000e+00, evaluated spans both sides

Falsification evidence (live runs)

Probe	Exit	Measured
use_mirror_merge = False	5	evaluated 252 ≠ 224 (doubled seam) — 5.1.2 and 4.5.11 identical
one centerline vert pulled off plane	4	27 ≠ 28 welded
mirror axis disabled	5	126 ≠ 224 — the half-car
modifier applied into data	3	datablock (224, 434, 212) ≠ (126, 231, 106)

Design note (in README): mutating a non-plane base vertex cannot break the
±X pairing — half ∪ mirror(half) is symmetric by construction — so the
check set targets the modifier, which is where realistic failures live.

Render / visual-pass log

Hero: candy-red dielectric-glass hatchback, default stage per
docs/VISUAL-STYLE.md (no deviation). 9 draft iterations: (1) glass
rule swallowed pillars → window band only; (2) headlamp strip painted the
nose white → lamps as separate mirrored boxes; (3) lamps floated → embedded
in the bumper band, bumper trim narrowed to the lower half; (4–6) windshield
rendered as a hot salmon slab — diagnosed as the paint-only center strips
(k=6/7) of the windshield slope catching the key, not the glass: the glass
rule now assigns the full windshield/rear-window slopes (steepest roof
rise/drop pairs) and the glass is dielectric, not metallic; (7) camera
stance. Pass 1 crops: no clipped highlights (windshield sheen ~0.75,
headlamps small and bright by design), lamps flush, contact shadow grounded.
Pass 2 falsification pair: mirror disabled renders half a car — halved
windshield/hood, one headlamp — wrong on sight. Pass 3: card webp reads as a
car in under a second; distinct from the soccer-ball card (subject, palette,
camera).

Contact sheet:
docs/gallery/contact-sheets/car-mirror-symmetry-contact-sheet.webp
(candidate TL; armature-bend TR; damped-track-aim BL; bmesh-gear BR).
Verdicts: stage darkness match ✓ · wedge warmth within family range ✓ ·
subject fill comparable ✓ · saturation: deep candy red, holds the lineup ✓ ·
thumbnail legibility ✓. Mean luminance 0.1372 vs set 0.2787 / 0.1722 /
0.3048 — darkest of the four, attributable to subject albedo (dark red
paint, black glass/tires), not underexposure: paint highlights and
windshield sit at full exposure without clipping. Same stage, same wedge —
not sortable as the odd one out.

Pass 4 (Playwright, live local server)

Card captured in situ at desktop (1280px) and ~390px mobile widths, plus the
detail page: image fills card media at both widths, teaches/WITNESSES copy
complete, car recognizable at both sizes. Generated HTML read
character-by-character: alt complete under the generator's 160-char cap
(first sentence sized to fit), no truncation, no markdown artifacts,
Unicode (∪, ±, −) intact, assert_alts_survive_dotted_paths passes. The two
new cards (this + soccer-ball-goldberg) sit side by side at the grid's end
and read as distinct works.

Hazards found while authoring (in code comments + README)

• Object.to_mesh_clear() takes no argument on current Blender
  (5.1.2 and 4.5.11) — passing the mesh raises TypeError.
• A base-mesh mutation cannot falsify mirror symmetry (always
  half ∪ mirror(half)) — falsification must target the modifier.
• Metallic glass mirrors the key light: the windshield renders as a hot
  salmon slab even though the material is "dark". Dielectric glass is the
  studio look.

Integration

Example dir + README + preview webp (1200×675) · hero webp (1280×720, true
RIFF via save_render — the documented img.save(filepath=..., quality=85)
path silently writes PNG bytes on these builds) · examples/gallery.json ·
.cursor-plugin/plugin.json · blender-smoke.yml step · README counts
37→38 + gallery row · ROADMAP pool drained (mirror-symmetry shipped; chess,
dice, bicycle directions remain) · gallery regenerated (38 detail pages).

Test plan

• Check-only, both binaries above: blender --background --python examples/car-mirror-symmetry/car_mirror_symmetry.py -- → exit 0
  (live-run locally on both, values above).
• Render path: exercised locally on 5.1.2 and 4.5.11 (EEVEE stills,
  cms_draft9*.png).
• CI: both blender-smoke matrix jobs (4.5, 5.1) must pass on this PR head.

Live-run vs inspection-only

• Live run: all checks (both binaries), all four falsification probes,
  both renders, webp conversion + magic-byte verification, gallery build,
  Playwright captures, contact-sheet luminance.
• Inspection only: CI runner behavior of the new smoke step; GitHub
  Pages deploy (path-filtered, post-merge).